Re: [PATCH 1/4] Add kswapd descriptor.

[KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>]

On Tue, 30 Nov 2010 12:40:16 -0800
Ying Han <yinghan@google.com> wrote:

> On Tue, Nov 30, 2010 at 12:54 AM, KAMEZAWA Hiroyuki
> <kamezawa.hiroyu@jp.fujitsu.com> wrote:
> > On Tue, 30 Nov 2010 17:27:10 +0900
> > KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:
> >
> >> On Tue, 30 Nov 2010 17:15:37 +0900
> >> Minchan Kim <minchan.kim@gmail.com> wrote:
> >>
> >> > Ideally, I hope we unify global and memcg of kswapd for easy
> >> > maintainance if it's not a big problem.
> >> > When we make patches about lru pages, we always have to consider what
> >> > I should do for memcg.
> >> > And when we review patches, we also should consider what the patch is
> >> > missing for memcg.
> >> > It makes maintainance cost big. Of course, if memcg maintainers is
> >> > involved with all patches, it's no problem as it is.
> >> >
> >> I know it's not. But thread control of kswapd will not have much merging point.
> >> And balance_pgdat() is fully replaced in patch/3. The effort for merging seems
> >> not big.
> 
> I intended to separate out the logic of per-memcg kswapd logics and
> not having it
> interfere with existing code. This should help for merging.
> 

yes.


> >>
> >
> > kswapd's balance_pgdat() is for following
> > A - reclaim pages within a node.
> > A - balancing zones in a pgdat.
> >
> > memcg's background reclaim needs followings.
> > A - reclaim pages within a memcg
> > A - reclaim pages from arbitrary zones, if it's fair, it's good.
> > A  A But it's not important from which zone the pages are reclaimed from.
> > A  A (I'm not sure we can select "the oldest" pages from divided LRU.)
> 
> The current implementation is simple, which it iterates all the nodes
> and reclaims pages from the per-memcg-per-zone LRU. As long as the
> wmarks is ok, the kswapd is done. Meanwhile, in order to not wasting
> cputime on "unreclaimable: nodes ( a node is unreclaimable if all the
> zones are unreclaimable), I used the nodemask to record that from the
> last scan, and the bit is reset as long as a page is returned back.
> This is a similar logic used in the global kswapd.
> 
> A potential improvement is to remember the last node we reclaimed
> from, and starting from the next node for the next kswapd wake_up.
> This avoids the case all the memcg kswapds are reclaiming from the
> small node ids on large numa machines.
> 
Yes, that's helpful.

> >
> > Then, merging will put 2 _very_ different functionalities into 1 function.
> 
> Agree.
> 
> >
> > So, I thought it's simpler to implement
> >
> > A 1. a victim node selector (This algorithm will never be in kswapd.)
> 
> Yeah, or round robin as I replied above ?
> 
I think it's good to have.

> > A 2. call _existing_ try_to_free_pages_mem_cgroup() with node local zonelist.
> > A Sharing is enough.
> 
> That will in turn use direct reclaim logic which has no notion of wmarks.
> 

 do {
	node = select_victim_node();
	do_try_to_free_pages_mem_cgroup(node);
	check watermark
 }

or If we need to check priority at el, your new balance_pgdat_mem_cgroup()
will be good.

> > kswapd stop/go routine may be able to be shared. But this patch itself seems not
> > very good to me.
> This looks feasible change, I will double check with it.

Thanks.

Regards,
-Kame

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>